src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  85     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  94     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  95     __m128           c6grid_00;
  96     real             *vdwgridparam;
  97     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  98     __m128           one_half = _mm_set1_ps(0.5);
  99     __m128           minus_one = _mm_set1_ps(-1.0);
 100     __m128i          ewitab;
 101     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 102     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 103     real             *ewtab;
 104     __m128           dummy_mask,cutoff_mask;
 105     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 106     __m128           one     = _mm_set1_ps(1.0);
 107     __m128           two     = _mm_set1_ps(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_ps(fr->ic->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124     vdwgridparam     = fr->ljpme_c6grid;
 125     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 126     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 127     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 128
 129     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 130     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 131     beta2            = _mm_mul_ps(beta,beta);
 132     beta3            = _mm_mul_ps(beta,beta2);
 133     ewtab            = fr->ic->tabq_coul_FDV0;
 134     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 135     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 136
 137     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 138     rcutoff_scalar   = fr->ic->rcoulomb;
 139     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 140     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 141
 142     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 143     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 144
 145     /* Avoid stupid compiler warnings */
 146     jnrA = jnrB = jnrC = jnrD = 0;
 147     j_coord_offsetA = 0;
 148     j_coord_offsetB = 0;
 149     j_coord_offsetC = 0;
 150     j_coord_offsetD = 0;
 151
 152     outeriter        = 0;
 153     inneriter        = 0;
 154
 155     for(iidx=0;iidx<4*DIM;iidx++)
 156     {
 157         scratch[iidx] = 0.0;
 158     }
 159
 160     /* Start outer loop over neighborlists */
 161     for(iidx=0; iidx<nri; iidx++)
 162     {
 163         /* Load shift vector for this list */
 164         i_shift_offset   = DIM*shiftidx[iidx];
 165
 166         /* Load limits for loop over neighbors */
 167         j_index_start    = jindex[iidx];
 168         j_index_end      = jindex[iidx+1];
 169
 170         /* Get outer coordinate index */
 171         inr              = iinr[iidx];
 172         i_coord_offset   = DIM*inr;
 173
 174         /* Load i particle coords and add shift vector */
 175         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 176
 177         fix0             = _mm_setzero_ps();
 178         fiy0             = _mm_setzero_ps();
 179         fiz0             = _mm_setzero_ps();
 180
 181         /* Load parameters for i particles */
 182         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 183         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 184
 185         /* Reset potential sums */
 186         velecsum         = _mm_setzero_ps();
 187         vvdwsum          = _mm_setzero_ps();
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 191         {
 192
 193             /* Get j neighbor index, and coordinate index */
 194             jnrA             = jjnr[jidx];
 195             jnrB             = jjnr[jidx+1];
 196             jnrC             = jjnr[jidx+2];
 197             jnrD             = jjnr[jidx+3];
 198             j_coord_offsetA  = DIM*jnrA;
 199             j_coord_offsetB  = DIM*jnrB;
 200             j_coord_offsetC  = DIM*jnrC;
 201             j_coord_offsetD  = DIM*jnrD;
 202
 203             /* load j atom coordinates */
 204             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 205                                               x+j_coord_offsetC,x+j_coord_offsetD,
 206                                               &jx0,&jy0,&jz0);
 207
 208             /* Calculate displacement vector */
 209             dx00             = _mm_sub_ps(ix0,jx0);
 210             dy00             = _mm_sub_ps(iy0,jy0);
 211             dz00             = _mm_sub_ps(iz0,jz0);
 212
 213             /* Calculate squared distance and things based on it */
 214             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 215
 216             rinv00           = avx128fma_invsqrt_f(rsq00);
 217
 218             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 219
 220             /* Load parameters for j particles */
 221             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 222                                                               charge+jnrC+0,charge+jnrD+0);
 223             vdwjidx0A        = 2*vdwtype[jnrA+0];
 224             vdwjidx0B        = 2*vdwtype[jnrB+0];
 225             vdwjidx0C        = 2*vdwtype[jnrC+0];
 226             vdwjidx0D        = 2*vdwtype[jnrD+0];
 227
 228             /**************************
 229              * CALCULATE INTERACTIONS *
 230              **************************/
 231
 232             if (gmx_mm_any_lt(rsq00,rcutoff2))
 233             {
 234
 235             r00              = _mm_mul_ps(rsq00,rinv00);
 236
 237             /* Compute parameters for interactions between i and j atoms */
 238             qq00             = _mm_mul_ps(iq0,jq0);
 239             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 240                                          vdwparam+vdwioffset0+vdwjidx0B,
 241                                          vdwparam+vdwioffset0+vdwjidx0C,
 242                                          vdwparam+vdwioffset0+vdwjidx0D,
 243                                          &c6_00,&c12_00);
 244
 245             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 246                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 247                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 248                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 249
 250             /* EWALD ELECTROSTATICS */
 251
 252             /* Analytical PME correction */
 253             zeta2            = _mm_mul_ps(beta2,rsq00);
 254             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 255             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 256             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 257             felec            = _mm_mul_ps(qq00,felec);
 258             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 259             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 260             velec            = _mm_mul_ps(qq00,velec);
 261
 262             /* Analytical LJ-PME */
 263             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 264             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 265             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 266             exponent         = avx128fma_exp_f(ewcljrsq);
 267             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 268             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 269             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 270             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 271             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 272             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 273                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 274             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 275             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 276
 277             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 278
 279             /* Update potential sum for this i atom from the interaction with this j atom. */
 280             velec            = _mm_and_ps(velec,cutoff_mask);
 281             velecsum         = _mm_add_ps(velecsum,velec);
 282             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 283             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 284
 285             fscal            = _mm_add_ps(felec,fvdw);
 286
 287             fscal            = _mm_and_ps(fscal,cutoff_mask);
 288
 289              /* Update vectorial force */
 290             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 291             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 292             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 293
 294             fjptrA             = f+j_coord_offsetA;
 295             fjptrB             = f+j_coord_offsetB;
 296             fjptrC             = f+j_coord_offsetC;
 297             fjptrD             = f+j_coord_offsetD;
 298             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 299                                                    _mm_mul_ps(dx00,fscal),
 300                                                    _mm_mul_ps(dy00,fscal),
 301                                                    _mm_mul_ps(dz00,fscal));
 302
 303             }
 304
 305             /* Inner loop uses 63 flops */
 306         }
 307
 308         if(jidx<j_index_end)
 309         {
 310
 311             /* Get j neighbor index, and coordinate index */
 312             jnrlistA         = jjnr[jidx];
 313             jnrlistB         = jjnr[jidx+1];
 314             jnrlistC         = jjnr[jidx+2];
 315             jnrlistD         = jjnr[jidx+3];
 316             /* Sign of each element will be negative for non-real atoms.
 317              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 318              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 319              */
 320             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 321             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 322             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 323             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 324             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 325             j_coord_offsetA  = DIM*jnrA;
 326             j_coord_offsetB  = DIM*jnrB;
 327             j_coord_offsetC  = DIM*jnrC;
 328             j_coord_offsetD  = DIM*jnrD;
 329
 330             /* load j atom coordinates */
 331             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 332                                               x+j_coord_offsetC,x+j_coord_offsetD,
 333                                               &jx0,&jy0,&jz0);
 334
 335             /* Calculate displacement vector */
 336             dx00             = _mm_sub_ps(ix0,jx0);
 337             dy00             = _mm_sub_ps(iy0,jy0);
 338             dz00             = _mm_sub_ps(iz0,jz0);
 339
 340             /* Calculate squared distance and things based on it */
 341             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 342
 343             rinv00           = avx128fma_invsqrt_f(rsq00);
 344
 345             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 346
 347             /* Load parameters for j particles */
 348             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 349                                                               charge+jnrC+0,charge+jnrD+0);
 350             vdwjidx0A        = 2*vdwtype[jnrA+0];
 351             vdwjidx0B        = 2*vdwtype[jnrB+0];
 352             vdwjidx0C        = 2*vdwtype[jnrC+0];
 353             vdwjidx0D        = 2*vdwtype[jnrD+0];
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             if (gmx_mm_any_lt(rsq00,rcutoff2))
 360             {
 361
 362             r00              = _mm_mul_ps(rsq00,rinv00);
 363             r00              = _mm_andnot_ps(dummy_mask,r00);
 364
 365             /* Compute parameters for interactions between i and j atoms */
 366             qq00             = _mm_mul_ps(iq0,jq0);
 367             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 368                                          vdwparam+vdwioffset0+vdwjidx0B,
 369                                          vdwparam+vdwioffset0+vdwjidx0C,
 370                                          vdwparam+vdwioffset0+vdwjidx0D,
 371                                          &c6_00,&c12_00);
 372
 373             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 374                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 375                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 376                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 377
 378             /* EWALD ELECTROSTATICS */
 379
 380             /* Analytical PME correction */
 381             zeta2            = _mm_mul_ps(beta2,rsq00);
 382             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 383             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 384             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 385             felec            = _mm_mul_ps(qq00,felec);
 386             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 387             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 388             velec            = _mm_mul_ps(qq00,velec);
 389
 390             /* Analytical LJ-PME */
 391             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 392             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 393             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 394             exponent         = avx128fma_exp_f(ewcljrsq);
 395             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 396             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 397             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 398             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 399             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 400             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 401                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 402             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 403             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 404
 405             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velec            = _mm_and_ps(velec,cutoff_mask);
 409             velec            = _mm_andnot_ps(dummy_mask,velec);
 410             velecsum         = _mm_add_ps(velecsum,velec);
 411             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 412             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 413             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 414
 415             fscal            = _mm_add_ps(felec,fvdw);
 416
 417             fscal            = _mm_and_ps(fscal,cutoff_mask);
 418
 419             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 420
 421              /* Update vectorial force */
 422             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 423             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 424             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 425
 426             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 427             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 428             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 429             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 430             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 431                                                    _mm_mul_ps(dx00,fscal),
 432                                                    _mm_mul_ps(dy00,fscal),
 433                                                    _mm_mul_ps(dz00,fscal));
 434
 435             }
 436
 437             /* Inner loop uses 64 flops */
 438         }
 439
 440         /* End of innermost loop */
 441
 442         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 443                                               f+i_coord_offset,fshift+i_shift_offset);
 444
 445         ggid                        = gid[iidx];
 446         /* Update potential energies */
 447         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 448         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 449
 450         /* Increment number of inner iterations */
 451         inneriter                  += j_index_end - j_index_start;
 452
 453         /* Outer loop uses 9 flops */
 454     }
 455
 456     /* Increment number of outer iterations */
 457     outeriter        += nri;
 458
 459     /* Update outer/inner flops */
 460
 461     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*64);
 462 }
 463 /*
 464  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_F_avx_128_fma_single
 465  * Electrostatics interaction: Ewald
 466  * VdW interaction:            LJEwald
 467  * Geometry:                   Particle-Particle
 468  * Calculate force/pot:        Force
 469  */
 470 void
 471 nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_F_avx_128_fma_single
 472                     (t_nblist                    * gmx_restrict       nlist,
 473                      rvec                        * gmx_restrict          xx,
 474                      rvec                        * gmx_restrict          ff,
 475                      struct t_forcerec           * gmx_restrict          fr,
 476                      t_mdatoms                   * gmx_restrict     mdatoms,
 477                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 478                      t_nrnb                      * gmx_restrict        nrnb)
 479 {
 480     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 481      * just 0 for non-waters.
 482      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 483      * jnr indices corresponding to data put in the four positions in the SIMD register.
 484      */
 485     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 486     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 487     int              jnrA,jnrB,jnrC,jnrD;
 488     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 489     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 490     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 491     real             rcutoff_scalar;
 492     real             *shiftvec,*fshift,*x,*f;
 493     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 494     real             scratch[4*DIM];
 495     __m128           fscal,rcutoff,rcutoff2,jidxall;
 496     int              vdwioffset0;
 497     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 498     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 499     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 500     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 501     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 502     real             *charge;
 503     int              nvdwtype;
 504     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 505     int              *vdwtype;
 506     real             *vdwparam;
 507     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 508     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 509     __m128           c6grid_00;
 510     real             *vdwgridparam;
 511     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 512     __m128           one_half = _mm_set1_ps(0.5);
 513     __m128           minus_one = _mm_set1_ps(-1.0);
 514     __m128i          ewitab;
 515     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 516     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 517     real             *ewtab;
 518     __m128           dummy_mask,cutoff_mask;
 519     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 520     __m128           one     = _mm_set1_ps(1.0);
 521     __m128           two     = _mm_set1_ps(2.0);
 522     x                = xx[0];
 523     f                = ff[0];
 524
 525     nri              = nlist->nri;
 526     iinr             = nlist->iinr;
 527     jindex           = nlist->jindex;
 528     jjnr             = nlist->jjnr;
 529     shiftidx         = nlist->shift;
 530     gid              = nlist->gid;
 531     shiftvec         = fr->shift_vec[0];
 532     fshift           = fr->fshift[0];
 533     facel            = _mm_set1_ps(fr->ic->epsfac);
 534     charge           = mdatoms->chargeA;
 535     nvdwtype         = fr->ntype;
 536     vdwparam         = fr->nbfp;
 537     vdwtype          = mdatoms->typeA;
 538     vdwgridparam     = fr->ljpme_c6grid;
 539     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 540     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 541     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 542
 543     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 544     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 545     beta2            = _mm_mul_ps(beta,beta);
 546     beta3            = _mm_mul_ps(beta,beta2);
 547     ewtab            = fr->ic->tabq_coul_F;
 548     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 549     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 550
 551     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 552     rcutoff_scalar   = fr->ic->rcoulomb;
 553     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 554     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 555
 556     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 557     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 558
 559     /* Avoid stupid compiler warnings */
 560     jnrA = jnrB = jnrC = jnrD = 0;
 561     j_coord_offsetA = 0;
 562     j_coord_offsetB = 0;
 563     j_coord_offsetC = 0;
 564     j_coord_offsetD = 0;
 565
 566     outeriter        = 0;
 567     inneriter        = 0;
 568
 569     for(iidx=0;iidx<4*DIM;iidx++)
 570     {
 571         scratch[iidx] = 0.0;
 572     }
 573
 574     /* Start outer loop over neighborlists */
 575     for(iidx=0; iidx<nri; iidx++)
 576     {
 577         /* Load shift vector for this list */
 578         i_shift_offset   = DIM*shiftidx[iidx];
 579
 580         /* Load limits for loop over neighbors */
 581         j_index_start    = jindex[iidx];
 582         j_index_end      = jindex[iidx+1];
 583
 584         /* Get outer coordinate index */
 585         inr              = iinr[iidx];
 586         i_coord_offset   = DIM*inr;
 587
 588         /* Load i particle coords and add shift vector */
 589         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 590
 591         fix0             = _mm_setzero_ps();
 592         fiy0             = _mm_setzero_ps();
 593         fiz0             = _mm_setzero_ps();
 594
 595         /* Load parameters for i particles */
 596         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 597         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 598
 599         /* Start inner kernel loop */
 600         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 601         {
 602
 603             /* Get j neighbor index, and coordinate index */
 604             jnrA             = jjnr[jidx];
 605             jnrB             = jjnr[jidx+1];
 606             jnrC             = jjnr[jidx+2];
 607             jnrD             = jjnr[jidx+3];
 608             j_coord_offsetA  = DIM*jnrA;
 609             j_coord_offsetB  = DIM*jnrB;
 610             j_coord_offsetC  = DIM*jnrC;
 611             j_coord_offsetD  = DIM*jnrD;
 612
 613             /* load j atom coordinates */
 614             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 615                                               x+j_coord_offsetC,x+j_coord_offsetD,
 616                                               &jx0,&jy0,&jz0);
 617
 618             /* Calculate displacement vector */
 619             dx00             = _mm_sub_ps(ix0,jx0);
 620             dy00             = _mm_sub_ps(iy0,jy0);
 621             dz00             = _mm_sub_ps(iz0,jz0);
 622
 623             /* Calculate squared distance and things based on it */
 624             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 625
 626             rinv00           = avx128fma_invsqrt_f(rsq00);
 627
 628             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 629
 630             /* Load parameters for j particles */
 631             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 632                                                               charge+jnrC+0,charge+jnrD+0);
 633             vdwjidx0A        = 2*vdwtype[jnrA+0];
 634             vdwjidx0B        = 2*vdwtype[jnrB+0];
 635             vdwjidx0C        = 2*vdwtype[jnrC+0];
 636             vdwjidx0D        = 2*vdwtype[jnrD+0];
 637
 638             /**************************
 639              * CALCULATE INTERACTIONS *
 640              **************************/
 641
 642             if (gmx_mm_any_lt(rsq00,rcutoff2))
 643             {
 644
 645             r00              = _mm_mul_ps(rsq00,rinv00);
 646
 647             /* Compute parameters for interactions between i and j atoms */
 648             qq00             = _mm_mul_ps(iq0,jq0);
 649             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 650                                          vdwparam+vdwioffset0+vdwjidx0B,
 651                                          vdwparam+vdwioffset0+vdwjidx0C,
 652                                          vdwparam+vdwioffset0+vdwjidx0D,
 653                                          &c6_00,&c12_00);
 654
 655             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 656                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 657                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 658                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 659
 660             /* EWALD ELECTROSTATICS */
 661
 662             /* Analytical PME correction */
 663             zeta2            = _mm_mul_ps(beta2,rsq00);
 664             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 665             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 666             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 667             felec            = _mm_mul_ps(qq00,felec);
 668
 669             /* Analytical LJ-PME */
 670             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 671             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 672             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 673             exponent         = avx128fma_exp_f(ewcljrsq);
 674             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 675             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 676             /* f6A = 6 * C6grid * (1 - poly) */
 677             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 678             /* f6B = C6grid * exponent * beta^6 */
 679             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 680             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 681             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 682
 683             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 684
 685             fscal            = _mm_add_ps(felec,fvdw);
 686
 687             fscal            = _mm_and_ps(fscal,cutoff_mask);
 688
 689              /* Update vectorial force */
 690             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 691             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 692             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 693
 694             fjptrA             = f+j_coord_offsetA;
 695             fjptrB             = f+j_coord_offsetB;
 696             fjptrC             = f+j_coord_offsetC;
 697             fjptrD             = f+j_coord_offsetD;
 698             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 699                                                    _mm_mul_ps(dx00,fscal),
 700                                                    _mm_mul_ps(dy00,fscal),
 701                                                    _mm_mul_ps(dz00,fscal));
 702
 703             }
 704
 705             /* Inner loop uses 52 flops */
 706         }
 707
 708         if(jidx<j_index_end)
 709         {
 710
 711             /* Get j neighbor index, and coordinate index */
 712             jnrlistA         = jjnr[jidx];
 713             jnrlistB         = jjnr[jidx+1];
 714             jnrlistC         = jjnr[jidx+2];
 715             jnrlistD         = jjnr[jidx+3];
 716             /* Sign of each element will be negative for non-real atoms.
 717              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 718              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 719              */
 720             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 721             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 722             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 723             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 724             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 725             j_coord_offsetA  = DIM*jnrA;
 726             j_coord_offsetB  = DIM*jnrB;
 727             j_coord_offsetC  = DIM*jnrC;
 728             j_coord_offsetD  = DIM*jnrD;
 729
 730             /* load j atom coordinates */
 731             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 732                                               x+j_coord_offsetC,x+j_coord_offsetD,
 733                                               &jx0,&jy0,&jz0);
 734
 735             /* Calculate displacement vector */
 736             dx00             = _mm_sub_ps(ix0,jx0);
 737             dy00             = _mm_sub_ps(iy0,jy0);
 738             dz00             = _mm_sub_ps(iz0,jz0);
 739
 740             /* Calculate squared distance and things based on it */
 741             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 742
 743             rinv00           = avx128fma_invsqrt_f(rsq00);
 744
 745             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 746
 747             /* Load parameters for j particles */
 748             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 749                                                               charge+jnrC+0,charge+jnrD+0);
 750             vdwjidx0A        = 2*vdwtype[jnrA+0];
 751             vdwjidx0B        = 2*vdwtype[jnrB+0];
 752             vdwjidx0C        = 2*vdwtype[jnrC+0];
 753             vdwjidx0D        = 2*vdwtype[jnrD+0];
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             if (gmx_mm_any_lt(rsq00,rcutoff2))
 760             {
 761
 762             r00              = _mm_mul_ps(rsq00,rinv00);
 763             r00              = _mm_andnot_ps(dummy_mask,r00);
 764
 765             /* Compute parameters for interactions between i and j atoms */
 766             qq00             = _mm_mul_ps(iq0,jq0);
 767             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 768                                          vdwparam+vdwioffset0+vdwjidx0B,
 769                                          vdwparam+vdwioffset0+vdwjidx0C,
 770                                          vdwparam+vdwioffset0+vdwjidx0D,
 771                                          &c6_00,&c12_00);
 772
 773             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 774                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 775                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 776                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 777
 778             /* EWALD ELECTROSTATICS */
 779
 780             /* Analytical PME correction */
 781             zeta2            = _mm_mul_ps(beta2,rsq00);
 782             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 783             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 784             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 785             felec            = _mm_mul_ps(qq00,felec);
 786
 787             /* Analytical LJ-PME */
 788             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 789             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 790             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 791             exponent         = avx128fma_exp_f(ewcljrsq);
 792             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 793             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 794             /* f6A = 6 * C6grid * (1 - poly) */
 795             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 796             /* f6B = C6grid * exponent * beta^6 */
 797             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 798             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 799             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 800
 801             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 802
 803             fscal            = _mm_add_ps(felec,fvdw);
 804
 805             fscal            = _mm_and_ps(fscal,cutoff_mask);
 806
 807             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 808
 809              /* Update vectorial force */
 810             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 811             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 812             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 813
 814             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 815             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 816             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 817             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 818             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 819                                                    _mm_mul_ps(dx00,fscal),
 820                                                    _mm_mul_ps(dy00,fscal),
 821                                                    _mm_mul_ps(dz00,fscal));
 822
 823             }
 824
 825             /* Inner loop uses 53 flops */
 826         }
 827
 828         /* End of innermost loop */
 829
 830         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 831                                               f+i_coord_offset,fshift+i_shift_offset);
 832
 833         /* Increment number of inner iterations */
 834         inneriter                  += j_index_end - j_index_start;
 835
 836         /* Outer loop uses 7 flops */
 837     }
 838
 839     /* Increment number of outer iterations */
 840     outeriter        += nri;
 841
 842     /* Update outer/inner flops */
 843
 844     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*53);
 845 }